Hi, my name is
Backend engineer working with Python and Java/Spring Boot. I've built Reva, a production AI agent, and a vectorless RAG pipeline. I like shipping systems that are fast, observable, and boring to operate.
I'm a backend engineer who enjoys the parts of software most people avoid: data models, query performance, caching layers, and making LLMs behave in production.
Recently my focus has been applied AI engineering: building RAG pipelines that skip the vector database entirely (BM25 + PostgreSQL full-text search), multi-agent systems that forecast inventory, and natural-language-to-SQL agents with self-correction and guardrails baked in.
Outside of work I grind competitive programming on LeetCode and Codeforces.
Production-ready RAG pipeline for catalog querying, with no vector database required. Hybrid retrieval combining PostgreSQL full-text search (tsvector + GIN) with BM25 reranking, a feedback loop that retunes term weights from user ratings, and Redis response caching.
Multi-agent AI system that forecasts product availability and restocking needs. Three specialized agents (demand forecasting, anomaly detection, catalog validation) run in parallel and an orchestrator synthesizes decisions. Every prediction is persisted with reasoning chains, confidence scores, and latency metrics.
Agentic natural-language-to-SQL system: two-stage intent extraction → SQL generation, self-correction with retry-on-error, a clarification agent for ambiguous questions, live schema introspection, and security-first guardrails (SELECT-only, keyword blocking, result limits).
Full-stack URL shortener with custom aliases, JWT auth, per-operation rate limiting, and per-day click analytics rendered with Recharts.
Multi-channel messaging platform unifying SMS, WhatsApp, email, and social media: Twilio webhooks, message scheduling via cron, role-based access, and an analytics dashboard.
Quiz platform decomposed into Java/Spring Boot microservices: service discovery, API gateway, and inter-service communication patterns.
More on GitHub →
I contribute to Toga, the Python-native, OS-native GUI toolkit behind BeeWare. Work there means writing the same feature eight times over, once per platform backend, and proving it on each one.
Added app.paths.desktop, .documents, .downloads and
.pictures to Toga's public API, implemented across
eight backends. On Windows the locations come from the OS, so OneDrive
folder redirection is honoured, and Downloads (absent from .NET's
SpecialFolder enum) is resolved through the Win32 known-folder API. On Linux
the lookup follows the freedesktop xdg-user-dirs resolution order. Platforms
with no user-accessible folders raise RuntimeError rather than inventing a path.
Shipped with testbed probes on every backend and 100% branch coverage on
core/src/toga/paths.py.
platformdirs for app paths
The maintainers opened this off the back of the work above: my PR made
platformdirs the basis for Linux paths, and the question is whether macOS and
Windows should converge on it too instead of keeping Toga's own duplicate implementation.
I'm auditing every path on both backends, comparing what platformdirs returns
against what Toga returns today, deciding which is actually correct under each platform's
conventions (and flagging where it is taste rather than standard), and mapping out a
migration path. The deliverable is a decision, not a patch.
Present
Building backend services and production AI systems, including Reva, a production AI agent, and a vectorless RAG pipeline.
05. What's next?
I'm open to interesting backend and AI engineering problems. Whether you have a question or just want to say hi, my inbox is open.
Say hello