I build production systems end to end — search pipelines that return in under a second, Go gateways that never starve a request, and RAG stacks written from the chunker up, down to the MCP server the model calls to search them itself.
US remote startup
DSA + system design
78-star OSS project
Hybrid RRF + cross-encoder
A year of production ownership — search infrastructure, two platform rebuilds, and the edge API underneath them.
Owned the search stack, an edge API, and two ground-up platform rebuilds for a high-traffic US firearms marketplace — four repositories, from the Elasticsearch query layer to the Cloudflare Worker in front of the database.
The product search layer in TypeScript — boosted multi-match queries with fuzzy matching, faceted aggregations across brand, caliber, capacity and retailer, and about two dozen filters that live in the URL. Sub-second over a large catalog, behind proxied routes that keep credentials off the client.
A ground-up SvelteKit rebuild of GunMade.com, from a shared component set that grew past 160 files to the technical SEO underneath it. Lazy loading and skeleton screens took Google PageSpeed from 60 to 90.
A full Svelte 5 rewrite that opened five new verticals — accessories, archery, knives, NFA and reloading — and a nationwide dealer locator that clusters tens of thousands of ATF-listed storefronts on a Mapbox map, also shipped as an embeddable widget.
The edge API every front end reads through: one Cloudflare Worker over two Hyperdrive bindings, serving product, blog, coupon and store data from a legacy Postgres and a newer Aurora cluster at once while traffic moved between them.
Shipped client-facing products across three frontend frameworks and a payments-backed Django backend — the stretch where I learned to pick up whichever framework a client had already chosen.
Built and deployed high-performance web applications with Next.js, Svelte and Angular for LoginRadius.com and EnrichlyHealth.com, increasing engagement and cutting page-load latency.
Built scalable Django backends with AWS S3 storage and Stripe payments on InPixr.com — raising transaction success rates by 25% and overall efficiency by 40%.
B.Tech, Electronics · Prayagraj, Uttar Pradesh
Systems I built end to end — including private repositories and client sites that are live in production.
A codebase RAG agent with nothing borrowed from a framework: three chunking strategies kept separate so they can be measured against each other, dense and lexical retrieval fused with Reciprocal Rank Fusion, then a cross-encoder rerank. The whole thing is exposed to the model as MCP tools over JSON-RPC, so it decides when to search instead of being handed context. Top-5 hit rate went from 90% to 100%.
# final round: no tools, so the model has to answer
tools = [] if last_round else [to_groq_tool(t) for t in mcp_tools]
message = groq.chat(messages, tools=tools).choices[0].message
for call in message.tool_calls or []: # the model asks,
output = await session.call_tool( # our code runs it
call.function.name, json.loads(call.function.arguments))
messages.append({"role": "tool",
"tool_call_id": call.id,
"content": output}) # …and loopTwo fixes merged into mloda — a plugin-based data-access framework for AI and ML.
A reader handed an explicit file hint it did not own would clear that hint and rescan the entire DataAccessCollection — silently binding to a different file than the caller named, and in edge cases letting a foreign file’s validation exception abort matching for a reader that should have declined cleanly. ReadFile and ReadDocument now return None and decline, so the caller’s disambiguation choice survives.
Four tox failures that appeared only on native Windows, all rooted in the backslash doing double duty as path separator and escape character: repr() doubling backslashes in a docs-fence allowlist lookup, a test fixture built with the wrong separator, an error message, and a path used unescaped as a regex pattern. Fixed with raw-string joins, as_posix() lookup keys, an OS-native fixture path, and re.escape().
What I reach for, grouped by where it sits in the system.
Hiring, contracting, or just want to compare notes on retrieval? Send a message.
I reply to everything within a day or two. If it’s about a role, mentioning the stack and whether it’s remote saves us both a round trip.
Open to full-time roles · Immediate joiner