Agent systemShipped
How I build with AI agents in the loop
AI writes a lot of my first drafts. Here is the workflow that keeps me — not the model — accountable for what ships.
Oct 2026
The lab
Where I work things out: agent systems with guardrails, AI experiments, research, and engineering notes from shipping real products. Some of it shipped; some of it is still an idea.
Anatomy of a guarded agent
The HR inbox agent in five steps. One is probabilistic; the other four are plain, tested code — that’s the point.
- 01Code
Check
Headers, spoofed reply-to and payroll-fraud patterns are caught before any model call.
- 02Model
Triage
The model returns strict JSON: a category, risk flags, confidence and a draft. It has no tools.
- 03Code
Decide
Eight independent holds must all pass before anything is sent automatically.
- 04Code
Act
Send, save a draft or escalate — recipients must equal the original sender exactly.
- 05Code
Report
Slack alerts without sensitive details, and a metadata-only audit log.
Notebook
Filter by kind of work.
10 shown
- Engineering noteShipped
Printing from the cloud to a printer on someone else’s Wi-Fi
Why the agent must dial out, how five printer transports differ, and why the bridge refuses to talk to public addresses.
Oct 2026
- Engineering noteShipped
Offline-first is mostly about conflicts
Caching is the easy part. Notes on field-level patches, stale-write refusal and private data when two phones edit the same book offline.
Oct 2026
- IdeaIdea
Physics-informed models beyond feature engineering
Next steps after the GSM paper: constraints in the loss, other fabric structures, and models that refuse to predict outside physical bounds.
Oct 2026
- Agent systemShipped
The model proposes, code disposes
A guardrail pattern for business agents: let the LLM classify and draft, but put every consequential rule in deterministic, tested code.
Sep 2026
- Engineering noteShipped
Reading a million-line file in a browser tab
Workers own the bytes, the main thread owns the pixels, and a scroll position is “line N plus an offset” — not a giant spacer.
Sep 2026
- ResearchPublished
One physical feature beat hyper-parameter tuning
In our GSM study, adding Pierce’s tightness factor did more for accuracy than tuning — and SHAP showed the model using it the way textile physics says it should.
Sep 2026
- Agent systemExploring
Trust ladders for business agents
Autonomy should be earned per task: observe, suggest, draft, act with approval, act and report. Notes from prototyping an agent control plane.
Jun 2026
- AI experimentShipped
Local Bengali speech-to-text
A browser recorder plus a local Whisper server that turns Bengali dictation into an editable transcript and exports it — with no API keys and no audio leaving the machine.
Jun 2026
- Interface experimentShipped
A booking widget in three themes
Guest booking form components — date and time pickers, a five-minute hold timer and nearby-slot suggestions — themed three ways for different restaurant brands.
May 2026
Research
Papers and research projects, with the methods and results written up.
Physics-Guided ML for Fabric GSM
Research paper accepted at IEEE ICACT 2026. On 5,648 production samples, an Extra Trees model reached a test RMSE of 4.10 g/m² (R² 0.952, MAPE under 0.5%) after adding a physics-derived tightness factor; SHAP and LIME confirmed the model learned textile physics.
Co-author (2nd of 6), presenting author
Smart Aquaponics with a Digital Twin
A course research paper (3rd of 4 authors) designing ESP32 sensor nodes for water quality, a Firebase telemetry backbone, a React monitoring and control dashboard, a digital twin for scenario testing and LightGBM-based anomaly prediction.
Co-author (3rd of 4)
Automated Weeding Robot
An Arduino Mega robot with ultrasonic obstacle sensing, GPS waypoint navigation, differential drive and a high-speed cutter. Removed 85% of weeds in the lab and 75% in the field.
Co-author (4th of 6)