I'm Vikash Maddi, a full stack developer at VMock. For four years I've built LLM agents used by 25,000+ people, the MCP servers and proxies they talk through, and the tracing and evals that catch them when they drift. On the side I make Ascend Daily, an AI learning coach.
Full stack developer since August 2022, working on AI resume products and the agent platform underneath them.
Built and run an LLM agent on Kubernetes for long-lived, stateful sessions, with an MCP server and HTTP proxy that expose internal tools behind schema-validated boundaries. Guardrails live in the schema, not in prompt instructions. Retries, timeouts and rate limits are handled around the model APIs.
25,000+people use it
Query plan analysis and index design took the hot API path from 400 ms to 30 ms. RAG-style retrieval over Elasticsearch with embeddings, chunking and relevance evaluation, and event-driven pipelines over AWS SQS.
13×faster, 400 to 30 ms
Span-level tracing with Arize Phoenix across the agent lifecycle, so a failure is attributable to the model, the prompt, the tool or the harness. Token, cost and latency are tracked per session. LLM-as-judge regression suites gate every model and prompt change before rollout, and a context-management layer with prompt caching cut inference cost on long sessions.
Everysession traced and priced
Zero-downtime migration of 90k+ accounts using dual writes, hash-based reconciliation and a staged rollout with rollback at every stage. Self-serve tooling for environment and tenant provisioning removed 30% of manual operational work.
99.9%data integrity across 90k+ accounts
Built on nights and weekends. Click a card to bring it forward.
Live at ascenddaily.in
Rise to the role you're aiming for.
An AI learning coach that turns any target role, from engineer to designer to UPSC, into a living plan you can actually finish. Structured paths break skills, courses and projects into steps, an AI coach builds the plan with you, and XP, levels and streaks reward finishing rather than planning.
FastAPI, Python, SQLAlchemy, LLM coaching, JWT auth
Developer tool for Claude Code, Python
Shrink the agent's context, and prove what it saved.
A drop-in HTTP proxy between Claude Code and the model. It scores conversation history by relevance, elides stale tool output behind stubs the agent can recall over MCP, and keeps the prompt prefix byte-identical between compactions so the KV cache survives. A SQLite ledger prices every request from billed usage and reports net dollars saved, cache-miss penalties included. Built harness first: 100% recall across five planted-fact scenarios.
Python, streaming HTTP proxy, MCP server, SQLite, Anthropic API, eval harness
Personal tool, Next.js on Vercel
Paste a job link. Review the application. Press submit.
Paste a Greenhouse, Lever or Ashby posting. The app reads it, writes a one-page resume tailored to the role from a fixed master profile, with every claim checked against the source so the model cannot invent experience, renders it to PDF with headless Chromium, and emails it along with the questions it could not answer. A local Playwright runner then fills the real form in a visible browser and waits. Nothing is submitted until I press the button.
Next.js, TypeScript, Workers AI, Playwright, Vercel Blob, Resend, Clerk
1 of 3. Use the arrows, the tabs, or swipe.
What I reach for, grouped by the job it does.
I'm open to product engineering and AI platform roles, anywhere. The quickest way to reach me is email.
vikashmaddi@gmail.comPrefer to call? Verify a work email and my number appears here.