a local coding agent on a small model - two experiments in how far a good harness can take it.

• v1: a terminal coding agent on qwen 2.5 with tool execution and controlled workspace access. it checks whether a "tests passed" claim is backed by an actual test runner.
• v2: rebuilt on one rule - never trust the model's claims, only its evidence. a task completes only when the required evidence exists: a file was really read, a test really ran, a command really exited 0.
• transactional edits: edit, test, keep or rollback as one operation. a bad edit never just sits there.
• honest sandboxing: docker isolation when a daemon is reachable, a restricted local backend when not - and every result records which one ran and what it actually guarantees.
• sqlite project memory with provenance, an eval harness where a change ships only with zero regressions, and 60 tests.

status: a work in progress, on purpose. single workspace, single model, no streaming - a harness for making one small model reliable, not a framework.

a data-cleaning pipeline where the ai proposes and a deterministic validator disposes.

• an llm recommends cleaning operations against a data profile. an independent validator approves or rejects each one. spark executes only what passed.
• experiments compare validated ai, unvalidated ai and plain spark. in the stub-provider run, quality issues went from 14 to 9.
• 480 tests, and a limitations section that says what it can't do: local sparksession, small fixtures, no real-llm run yet.

status: research prototype with an ieee-format paper. the next experiment is the model-comparison matrix: llm vs slm, cloud vs local.

messaging for when the network is the problem. offline-first chat with signed identities and device-to-device relay.

• ed25519 identities: every device signs who it is and every message it sends. the receive path verifies the signature; it never trusts the peer.
• local-first storage: sqlite holds message history and a forward queue, so nothing depends on a connection existing.
• hop-limited relay: messages travel device to device with dedup and a ttl, so the mesh floods politely.
• transports: ble scan, connect and chunked transfer on the client side; wi-fi direct discovery is halfway there. the "chain" in the name is a per-device hash chain, not a blockchain.

status: prototype. a polished next.js demo walks the chat, nearby-peers, broadcast and sos flows; the react native core (~1.8k lines) carries the real architecture, with jest tests over mocked native modules. not a proven multi-device mesh yet.

a smart-city traffic operations dashboard - a command center for intersections, incidents and predictions.

• multi-page command center: operations, a traffic map (mapbox, with a canvas fallback), analytics, incidents, predictions and simulations.
• product engineering over plumbing: reusable components, stateful filtering and selection, six local api routes standing in for a backend.
• the map degrades gracefully: no mapbox token, no problem - the canvas draws the grid anyway.

status: prototype with simulated data. no live feed, no model, no database. the interesting work is the information architecture and the ui engineering.

the execution engine of a team-built agent workflow platform - the part that actually runs the workflows.

• hasura/nhost serverless functions in typescript over postgres: manual triggers and webhook triggers, both feeding one shared step executor.
• webhook security done right: an external caller needs a verified per-trigger secret, and sensitive steps run only because creating that trigger already required the owner role.
• authorization is derived server-side - the org is read from the workflow row, never from client input.
• quota admission, retries, and step handlers for db writes and notifications.

status: my ownership area on a divided build: the functions layer. no readme yet; the code carries it.

a management platform for kanwar yatra pilgrims, built at nic, meity.

• python app: pilgrim registration, groups, routes and checkpoints, check-ins and sos handling.
• role-based access, and qr identity cards for pilgrims.

status: demo data only. checkpoint/map and sos screens are still to build.

a strictly read-only system health collector for windows 11. it reports what it measured, and what it could not measure.

• one shot, ordinary user privileges: os, cpu, memory, disk, network, plus a schema version and a utc timestamp.
• deliberately omits hostnames, serial numbers and account names - a health report does not need to identify the machine.
• no diagnosis, no repair, no llm in the collector. honest measurement is the layer; interpretation sits on top of it separately.

status: small on purpose. the scope discipline is the point.

local-first context infrastructure for ai systems - one project's decisions and reasoning, consistent across chatgpt, claude and whatever model comes next.

• born from re-explaining projects to every model, and paying for it in credits and contradictions.
• docs-first by design: the ontology (what exists in a user's world) and the lifecycle (how beliefs get created, revised and retired) come before any storage or framework choice.

status: phase 1 is deliberately docs only. the backend is near-empty on purpose - the cognitive model is the product right now.