arXiv:2608.04804
Built the system and ran the experiments: the training pipeline for the 7B searcher, the verification gate that strips unsupported claims, the cost-aware router, and the evaluation harness.
Adithyan K
I build AI systems, and the tools that prove they work.
I trained a 7B model that reads a codebase before an agent tries to fix it, built the retrieval layer inside our coding agent, and wrote the evaluation platform that tells us whether any of it actually helped.
Two papers came out of that this year. Before the research work I shipped the sandbox and deploy infrastructure behind SuperAGI's app builder, and the backend for AI agents that join live meetings with voice and avatar. On the side I co-founded Lifie, where we build voice and chat AI for businesses.
Research
arXiv:2606.22417
Built the evaluation platform and ran a three-arm, three-seed ablation over 819 agent runs, including the leak-audit protocol and a localisation metric designed not to flatter our own system.
Selected work06
All projectsSuperScout· SuperAGI
Cost-aware routing for coding agents. A 7B model I trained scouts the repository first, and its verified findings decide which frontier model is worth paying for. I built the training pipeline and the evals that benchmarked four frontier models head to head on SWE-bench Pro, where the system matches the best single model at about a fifth the cost per solve.
- 1/5
- the cost per solve
- $86
- to train the searcher end to end
- 3,000+
- agent eval episodes run
SuperCoder· SuperAGI
SuperAGI's open-source, local-first coding agent, which searches a codebase by meaning and call graph rather than grep alone. I built the Rust client for its context engine, and the eval platform behind our paper, running large-scale agentic evals to test whether the index actually helps.
- 10
- languages parsed
- 819
- agent eval runs in the ablation
Digital Employees· SuperAGI
Autonomous AI teammates that join a live meeting with voice and avatar, run the sales demo, answer questions, and drive the slides themselves. I built the backend: the agent loop, the speech pipeline, and the meeting integration.
- 4,000+
- client meetings joined in production
Vibe Coder· SuperAGI
Describe an app, watch it get built, and deploy it to its own subdomain. I owned the execution and deployment side: the sandboxed runtime the agent writes code inside, and the pipeline that ships the result.
- 450+
- apps generated and served
Lifie
Voice and chat AI for businesses, across outbound engagement and inbound support. Co-founded it and own the architecture: the intelligence layer, retrieval grounding, and guardrails.
cc-vault
A terminal UI for browsing and managing Claude Code sessions. Search across every conversation, preview them as rendered markdown, and resume any one of them in place.
Experience
Training models, building the systems around them, and publishing the results, across coding agents, retrieval, and agent infrastructure.
- AI Research Engineer
- Trained a 7B repository-search model end to end, covering data curation, a custom packing pipeline, and the training run, then built the verification gate and cost-aware router around it. Published as SuperScout.
- Built the evaluation platform behind both papers: sandboxed run orchestration, leak detection, and cost accounting across 800+ agent runs.
- Second author on two arXiv preprints, having built the systems and run the experiments for both.
- AI Engineer, Intern
- Built the Rust client for SuperCoder’s context engine: repository parsing, structure-aware chunking, and a resumable content-addressed sync protocol that runs entirely on the developer’s machine.
- Shipped the sandbox runtime and deployment pipeline for Vibe Coder, SuperAGI’s AI app builder.
- Built the backend for Digital Employees: the agent loop, speech pipeline, and meeting integration behind voice-and-avatar agents that run live sales demos.
Voice and chat AI for businesses, across outbound engagement and inbound support.
- Co-Founder & Technical Lead
- Own the architecture end to end: LLM orchestration, context management, retrieval grounding, and guardrails.
- Built cross-platform interoperability on MCP.
- AI Product Developer, Intern
- Built agent-based automation for lead intelligence and outreach on QpiAI’s Agent Hive platform.
- Delivered a sales support agent adopted by the internal team.
- Full Stack Developer, Intern
- Built a license management platform with tiered access control, expiry alerts, and CSV export.
Beyond code
Football
The one thing outside work I take seriously. I would rather be playing than watching, but I do plenty of both.
Cooking
Currently more ambition than ability. I like the idea of being good at it enough to keep working at it.
Travel
Random trips, planned as little as possible. New places, new food, new people, mostly figured out after landing.
Contact
Open to research and applied AI roles.
Happy to talk about anything on this page, the papers especially.