From field problems
to working software.

01 / SE Command Center

Keep the next move clear on every opportunity

The problem

SE work (discovery, demos, RFPs, security questionnaires, handoffs) is spread across tools, so blockers, owners, and deadlines get lost.

The workflow

Opportunities move by stage from discovery to handoff. My Focus shows what needs attention, dependency tracking shows who you're waiting on, and calendar views turn deadlines into a plan. An Impact view shows the pipeline supported per SE hour. Suggest next step has Claude draft the next action, owner, and due date; nothing changes until the SE saves it.

Try it yourself

Explore a fictional SE workspace with 26 sample opportunities, RFP and InfoSec deadlines, and a guided tour.

Launch Interactive Demo ↗

Fictional data, no sign-in. Changes stay in your browser tab.

How I built it

Built in a few hours over about 60 iterations with ChatGPT and Claude, modeled on how I ran opportunities as an SE. Next.js and TypeScript on Vercel; my own workspace is behind a server-checked password.

Technologies & design

02 / SightFlow

One project record from survey to handoff

The problem

Security projects lose time and accuracy when field notes, device layouts, estimates, and installation instructions are managed separately.

The workflow

A guided site survey by area feeds a design workspace for equipment, labor, and BOM, then produces sales and installer handoff documents.

How I built it

Built from years of walking sites, over 149 iterations with ChatGPT, Claude, and Gemini. A static JavaScript app on Vercel with Supabase sign-in and project storage, JSON import and export, and built-in self-checks.

What's inside

A labor engine, cable takeoffs by type, IDF assignments with Mercury controller capacity validation, Verkada license and PoE rules, automatic NVR balancing, and lift scope review.

Demo project

A fictional 16-area hospital with 271 device records, 43 reader locations, 26 cameras, and about 41,000 feet of cable, producing a complete work order and sales BOM.

Technologies & design

03 / Ask Ky

Turn a static résumé into a conversation

The problem

Recruiters have specific questions a résumé can't answer at a glance: has he done this, with which tools, and how does it fit our role? Scrolling a long portfolio to find out takes time they don't have.

The solution

Ask Ky turns the portfolio into a conversation. Visitors ask in their own words and get a short answer with the sources it came from.

• Typed or spoken: tap the microphone and the browser's speech recognition sends the question.
• Suggested questions help a visitor get started, and follow-ups like "tell me more" keep the context.
• Check fit for a role: paste a job description and see how my background maps to each requirement.
• Next steps: download my resume or leave contact details without leaving the chat.
• Stays on topic: off-topic or unsupported questions get a polite boundary, not an invented answer.

How it works

A Vercel serverless function keeps the API key on the server, limits input, and rate-limits. Keyword retrieval picks the most relevant of about 40 knowledge-base sections, Claude answers only from them, and the server drops any citation it didn't send.

Claude Haiku 4.5 replies in JSON: the answer, the sections it cited, and whether it really answered. The browser shows the answer as plain text, never HTML.

Check fit for a role uses the same function but sends every section, so a match isn't limited by what retrieval picked. Visitor questions are logged, so I can see what recruiters actually ask and fill gaps in the knowledge base.

How I built it

I curated the knowledge base: about 40 sections on my experience, certifications, and every project, updated whenever a project changes. I chose keyword retrieval with synonyms over embeddings because it's simple, free to run, and easy to debug when an answer misses.

Two test scripts keep it honest. A retrieval test runs 106 realistic recruiter questions, many worded differently from the knowledge base on purpose, through the real retrieval code at no API cost: 105 find a section that answers them. A red-team script attacks the live endpoint with prompt injections and jailbreaks (see how I hardened it).

Technologies & design

04 / PowerQuote

A clearer path from first contact to payment

The problem

Small electrical contractors can lose track of customer requests when calls, texts, quotes, and follow-ups live in different places.

The workflow

PowerQuote keeps every customer request in one place, from the first message to the paid invoice.

• Leads and conversations in one inbox, with urgent requests flagged for a person to follow up.
• Quotes built line by line with parts, labor, and tax, sent to the customer by secure link.
• Invoices and customer feedback close out each job.
• Company dashboards surface the next action; a platform view shows activity and quote progress across companies.
• English and Spanish, with role-based access for each company's team.

Try it yourself

Explore a fictional electrical contractor account with sample leads, quotes, and invoices, plus a guided tour.

Launch Interactive Demo ↗

Fictional data only; external messaging is disabled.

Technologies & design

Open PowerQuote ↗

05 / PicTalk

Talk through the site walk instead of typing it up later

The problem

On a site walk, contractors take dozens of photos and quick notes. By the time they're back at the office, the context is gone, and someone has to rebuild it from memory before the estimate can start. Details like model numbers and what the customer asked for get lost.

The experience

Take a photo, tap to talk, save the stop. PicTalk pairs each photo with a voice note and its transcript, and groups stops by job.

• Works with no signal: installs like an app; stops wait on the phone, then upload and transcribe automatically.
• Describe photo: Claude reads the photo, copies model and serial numbers exactly, and notes visible condition, without guessing.
• Wrap-up notes: record field notes and customer comments before ending a job. On Android, words appear as you speak, and you can pause and fix the text.
• PDF report: each photo with its transcript and the time it was taken, shared straight from the phone.
• AI job summary: ranked action items linked to their stops, marked as AI-written; only an approved summary goes in the PDF report.
• Piccolo: turns a finished job into a work order, parts list, and quote. Every line links to its photo; the estimator edits, prices, and finalizes. Field techs never see prices.
• Help when stuck: tap ? and ask PicTalk's help chat, by typing or by voice.

Piccolo finalizes a numbered version that can't be changed and exports it as a PDF, CSV, or JSON. A company can invite estimators, field techs, and installers, and an admin console tracks each job from captured to quoted, won, and installed. Each final records what the estimator changed from the AI's draft, so later drafts reuse the company's own wording, part numbers, and prices.

How it works

• Stops queue in IndexedDB, upload to Firebase Storage, then write one Firestore record, so a record never points to missing files.
• Cloud Functions send audio to Deepgram with a security-trade vocabulary.
• Live words: a Cloud Function trades the server-held Deepgram key for a 30-second token, so the key never reaches the phone. Words show gray while Deepgram decides and white once final, and pausing closes the connection so silence isn't billed. If the connection drops, the full recording is transcribed instead, and typed text is never overwritten.
• Describe photo: the tap just marks the stop, so it works offline. A Cloud Function claims the request so it runs once, waits for the voice note's transcript, shrinks the photo, and sends both to Claude. A delete during the run wins.
• Claude returns JSON in a fixed schema. Code checks every line points to a real stop, marks unconfirmed part numbers Verify, and never invents prices.
• No sign-up to start: everyone gets an anonymous account, and signing in with Google or an email link keeps that same account, so nothing has to move.
• Security rules limit every record to its owner or company role, tested against Firebase emulators.
• AI summaries and photo descriptions are capped per job and per account per day, to keep costs predictable.
• Engineering Mode shows whether a slow stop was the phone's signal, the upload, or the speech-to-text service. It's off by default and shows each user only their own data.
• The Evaluation Lab cuts the signal at the worst moments. The first run caught three bugs; after the fixes, all 21 runs pass (October 4, 2026).

Try it yourself

Demo: designed for phones, up to 10 stops per job. Voice recordings are deleted after 5 days.

Where it's going

Next: quote approval by secure link, email delivery, and export into estimating tools and CRMs.

How I built it

I built PicTalk with Claude Code, Anthropic's AI coding agent, and learned Firebase, Node.js, and Git along the way. I started with the site walk itself (photo, voice, save) and made it work with no signal before adding any AI.

Then I made the hard parts measurable. Engineering Mode shows where the time goes on each stop, and the Evaluation Lab drives the real app in a phone-sized browser while cutting the signal. Its first run found that losing signal mid-sentence kept only the words heard before the drop. Now a lost signal transcribes the whole recording, and all 21 runs pass. I also tested on a real Android phone with made-up customers.

Technologies & design

06 / NightAgent

Stay with a service call from the first ring to a confirmed fix

The problem

After-hours calls reach an answering service that can only take a message. Emergencies wait with routine requests, and nobody checks that the fix held.

The experience

Sam, the AI voice agent, answers the night desk for a demo security integrator:

• Checks safety first, finds the account, confirms the caller's details, and opens a ticket on a live dispatch board. Emergencies page the on-call tech.
• Takes messages for any department and hands billing and sales calls to two specialist AI agents, Jordan and Riley.
• After the simulated repair, NightAgent calls the customer back. The answer closes, reopens, or escalates the ticket, or creates a sales lead.
• Built for non-technical users: one button, plain-English status, voice or typing.
• Options for slower speech and bigger text, one of five voices on each call, and it installs on a phone's home screen like an app.

How it works

ElevenLabs Agents handles speech and voice in real time. Sam calls tools on a Python FastAPI backend (look up customer, create ticket, page tech, take message), authenticated with a shared secret. The model suggests a priority, but rules in code decide: it can raise priority but never downgrade an emergency, and a ticket lifecycle in code prevents skipped stages. Every step is saved to Supabase, and a live Business Impact panel counts calls, fixes, and leads, with no invented revenue.

The follow-up call is a second ElevenLabs agent. A single-use pass ties it to the ticket, the agent reports what the customer said in one structured tool call, and code decides the outcome:

• Confirmed fix: the ticket closes.
• Problem came back: the ticket reopens and the priority is re-checked.
• Repair never worked: it escalates to the service manager.
• Upgrade interest: a sales opportunity for the account executive, value TBD unless the customer names a budget.
• Unclear: a person follows up instead of the AI guessing.

Every call also gets a call check: up to nine checks, mostly computed from the call's records, with two graded by ElevenLabs from the transcript: no unsupported promises, and the caller's details confirmed before acting.

How I engineered it

Problem	After-hours calls need the right urgency. A missed emergency can leave a site unsecured overnight; a false one sends a technician out at night for nothing.
Constraint	The AI can't be the final judge of priority, and the demo is open to anyone, so it has to be safe to leave running.
Decision	Rules in code set the final priority and decide what happens after each call. Sam suggests a priority; the code can accept a higher one but never lowers an emergency.
Hard part	The same problem arrives in different words ("our gate is stuck open", "the gate still sticks"). Some calls sound odd but aren't a problem yet, like a cat staring at the electrical panel; others are worse than they sound, like a burning smell from the alarm panel.
What I tried first	Putting the rules only in Sam's instructions. With this model that wasn't reliable: Sam sometimes skipped reading back the caller's name and number. Moving the next step into each tool's answer, so the server tells Sam exactly what to do, fixed it.
Evidence	21 regression tests run 3 times each against the live agents, including five prompt-injection attempts, plus 24 voice tests with real audio in noise and over a phone line. Every fix is re-tested against every test.
Limits	Real-call samples are still small, voice tests use recorded voices and a simulated phone line, and dispatch is simulated.

Live results and the fixes log: Evaluation Lab ↗ · Code: GitHub ↗

Guardrails and testing

• Never asks for alarm codes, explains bypasses, or invents account details, arrival times, or prices.
• Calls are time-limited, rate-capped, and accepted only from approved sites.
• Engineering Mode times every tool call; the Evaluation Lab scores tests and real calls separately, with sample sizes.
• Voice tests: 18 of 18 phone numbers heard exactly, about 1.7 seconds from the caller's last word to Sam's reply.
• Failure injection replays known failures (a duplicate call, a paging outage) through the real server code in a sandbox.
• A fixes log lists 14 problems found this way: why each happened, what changed, and whether the fix still holds in the latest runs.

Jargon Bench: does speech-to-text hear security jargon?

Question	How often do Deepgram and ElevenLabs hear trade terms like Verkada, Wiegand, NVR, and PoE correctly, and does keyterm boosting help?
Method	44 recorded field sentences, each sent to Deepgram nova-3 and ElevenLabs scribe_v2 three times, with and without boosting. The scorer and accepted spellings were frozen before the first run.
Result	Boosting raised accuracy from 62.1% to 89.7% (Deepgram) and 79.3% to 95.4% (ElevenLabs), adding 9 to 21 ms. Unboosted, "Wiegand" became "weekend"; boosted, it sometimes wrote terms that weren't said.
Limits	One speaker and 44 clips, so the results are directional, not definitive.

Full results: Scorecard ↗ · One-page readout (PDF) ↗ · Code and method: GitHub ↗

Try it yourself

1. Press Start call (or type) and report a problem, for example: "This is Sunset Dental, 310-555-0142. Our back door won't lock."
2. Sam opens a ticket. Hang up and watch it run: dispatched, en route, onsite, completed.
3. When the button rings, tap Answer and tell Sam how the repair went, then see the outcome and Business Impact update.

Or ask about an invoice to meet Jordan, ask about adding cameras to meet Riley, or pick a scenario in Demo Mode to skip the first call.

Made-up demo data; please don't share real names, numbers, or alarm codes. Conversations are processed and recorded by ElevenLabs. Later steps and staff are simulated.

Technologies & design

07 / Bluey

Get better results from AI without learning prompt engineering

The problem

Many people don't think of themselves as AI users. They don't know what to ask or how much context to give.

The experience

Bluey is a friendly blue orb you talk to in plain language. A clear request gets done right away; a vague one gets one question first. When there's a better way, he finishes the task, then mentions it.

• Steer buttons under every draft (Shorter, Warmer, More specific) teach by doing: each sends a plain message in your own chat.
• Scam check: paste a text, email, or screenshot for a clear verdict, the red flags, and what to do next.
• Practice talks: rehearse a hard conversation, then get kind feedback.
• Reads PDF, Word, and CSV files, speaks his replies, creates PDF, Word, and Excel files, and shows openly licensed photos from Wikipedia.
• Easier to use: bigger text, a slower voice, and simpler words.
• A real character with a backstory and gentle humor, so he never sounds the same twice.

How it works

A Vercel serverless function keeps the OpenAI key on the server and calls the Responses API with a strict JSON schema: the reply, a mood for the orb, and a hidden "brain" record (mode, goal, and an initiative level from 0 to 4). Every OpenAI endpoint has input limits and hourly and daily caps.

Memory and optional Google sign-in: Bluey remembers your name and goals with an Undo on every save, never keeps sensitive data, and syncs across devices through Supabase with row-level security. Without sign-in, everything stays on your device.

Brain Lab: 27 test conversations (39 release-gate tests on the beta branch) run through the real chat endpoint, scoring task success, initiative, question discipline, and character. They caught "Make it shorter" dropping every fact from an email; after the fix, revisions kept every fact in 10 of 10 runs.

How I built it

Designed over a week with ChatGPT: who Bluey is for, when he acts or asks, his personality, and the memory database. Since October 6, 2026, built with Claude Code, with every feature gated on Brain Lab or browser tests, including on a real iPhone.

Technologies & design

Alpha in active testing. Chat history stays in your browser; please don't share sensitive information.

08 / PC Medic

Give everyday Windows users a technician who looks before answering

The problem

When a computer acts up, people get generic advice written for someone else's computer. A good technician looks at this computer first. PC Medic makes that first look available to anyone.

The experience

PC Medic installs like any Windows program. A three-second read-only check gives a one-line verdict and five status tiles: speed, storage, battery, updates, and stability.

• Home mode speaks plain English; Technician mode shows raw values, error codes, and every Windows tool.
• Ask in your own words: "can I add more RAM?" reads the actual slots and modules.
• Investigations: "it keeps freezing" runs the right checks, correlates Windows logs, crashes, and drive health, ranks the likely causes, and offers the next test.
• Guided fixes measure before and after. PC Medic never changes anything itself.
• Deep checks: security posture, blue screens with the likely driver, Windows Update errors explained, and a read-only scan as administrator.
• Technician reports ready to print and sign.

How it works

A local Python server runs read-only PowerShell and performance-counter checks. It answers on the PC when it can; otherwise Claude (with read-only tools and cited web search) or ChatGPT helps, only after the user approves, with serial numbers, names, and addresses stripped first. Every answer says who answered it.

It's secured like any local web server: bound to 127.0.0.1, DNS-rebinding and CSRF protection, per-account access on shared PCs, and API keys in Windows Credential Manager. A per-user installer (PyInstaller and Inno Setup) needs no admin rights and updates from GitHub Releases.

How I built it

Built with ChatGPT through version 1.16, then with Claude Code from October 10, 2026, which added investigations, guided troubleshooting, Deep checks, the redesign, and the installer. Over 200 automated tests mock every AI call, and install and upgrade were tested on real hardware. Not done yet: code signing and testing on more computers.

The deep scan needed administrator rights without running the whole app as administrator. PC Medic starts a second copy of itself through the Windows UAC prompt that runs only fixed read-only steps, writes only its own result file, and exits. A real-machine check also caught another AI's mistake: it had said my Lenovo laptop's RAM was soldered, and PC Medic correctly read removable SO-DIMM modules.

Technologies & design

For Windows 10 and 11 laptops and desktops. The installer isn't code-signed yet, so Windows may show "Windows protected your PC": choose More info, then Run anyway. Works offline; AI answers need your own Anthropic or OpenAI key.

09 / PropertyIQ

Let a private AI explain the numbers without ever doing the math

The problem

Lending teams want AI answers, but their data can't go to a hosted model, and language models are unreliable at math. PropertyIQ runs the model next to the data, and the model explains results but never calculates them.

The experience

An analyst sees 55 synthetic properties across office, retail, industrial, and multifamily. Rule checks flag 4 as underperforming: occupancy below 85%, NOI down 10% or more year over year, or NOI at or below 90% of budget. Ask "Why was Westlake flagged?" and the answer links to its inputs, formulas, and record IDs.

It has grown into a lending workbench:

• Loan screening: DSCR, LTV, debt yield, and stress tests, exported to PDF and to Excel with live formulas.
• Document intake: rent rolls and T-12s from Excel, CSV, or scanned PDFs via local OCR, reviewed before use.
• Red flags: 13 document checks, a market check, and California rules, each with its evidence and a suggested action.
• Underwriting comparison: the analyst approves each adjustment with a reason code, and a summary built in code reports the result, for example DSCR moving from 1.58 to 1.00, failing the 1.25 minimum.
• Debt dashboard: 46 synthetic loans, a maturity wall, a DSCR × LTV risk map, and a covenant watchlist.
• AI Engineering Lab: request traces and a model on/off comparison.
• Admin: a live security checklist, an internet lockdown switch, and backups restored and checked row by row.

How it works

Application code calculates NOI, trailing-12-month totals, variance, and occupancy from PostgreSQL. A code router picks a fixed, access-scoped query, and only verified facts reach a local Qwen3 model on Ollama. Every number it writes is checked; on a mismatch, a summary built in code is shown instead.

Following one question: the analyst asks "Summarize our underperforming assets." The API queries only the data that analyst may see, code computes the figures and rule results, the model writes the summary from those facts alone, and the interface shows it next to the formulas and source records.

Four roles with fund-scoped access enforced in SQL; a hidden loan returns the same "not found" as one that doesn't exist. It's a prototype on synthetic data; production hardening is still to do.

How I built it

I built PropertyIQ with Claude Code on one Windows 11 laptop: Next.js 16 and React 19 for the interface, a FastAPI (Python) API, PostgreSQL 17 in Docker, and a 4-billion-parameter Qwen3 model served by Ollama. Documents are read with RapidOCR and pdfplumber, and exports use ReportLab and openpyxl, all on the same machine.

One rule shaped every module: code does the arithmetic and makes the decisions, and the model only explains. The loan tools use no AI for any number, each Excel export is recalculated in OpenOffice during testing and must match the app, and traces never store the question or answer text.

Try it yourself

The real app runs locally, so a browser simulator recreates its main screens with synthetic data.

Launch simulator demo ↗

Synthetic data only, no sign-in. It runs in your browser and nothing is saved or sent. The chat is scripted; in the real app, answers come from a local model.

Technologies & design

10 / WeldMate

Turn welding reference material into an interactive learning tool

The challenge

Welding students and field users often move between calculators, formulas, notes, procedure information, and practice material. WeldMate brings those activities into one focused mobile experience.

The experience

Calculators (heat input, fillet, V-groove, carbon equivalent), an eight-section learning path, quizzes, weld-symbol and inspection practice, and WPS practice cards, saved in the browser.

How I built it

Built with Google Gemini as a single mobile-first tool.

Technologies & design

Open WeldMate ↗

11 / Last Light Outpost

Build a base that can survive the next wave

The challenge

Make an accessible browser game where the player can see the consequences of each choice as enemy waves grow stronger.

The experience

Defend the last working generator for miles through escalating zombie waves. An active prototype I keep refining through play.

• Fight, collect, build: zombies drop coins the player must pick up, then spend on walls, turrets, carrying capacity, facilities, and crew.
• Choose a run: commanders with different abilities, several starting bases, and a full 56-wave survival mode.
• AI Director: Fair adapts pressure to the player's defenses and eases off near defeat; Ruthless never lets up.
• Forward operations: short missions with one clear goal at a time, like escorting a truck or landing a plane.
• Daily challenges, bonus missions, and online leaderboards.

How I built it

The whole game is one HTML file of plain JavaScript, with no engine or build step. Progress saves in the browser and syncs to Supabase in its own game schema. Instead of passwords, each device gets a random ID and secret, and a temporary transfer code moves progress to another device. A service worker makes it installable and playable offline, with keyboard, mouse, and touch controls.

Technologies & design

Play Last Light Outpost ↗