Skip to content
RAHUL.HONGEKAR

RahulHongekar

AI / Software Engineer

Building AI agents, intelligent architectures, and the systems that bring them to life.

Explore workLinkedIn ↗
Scroll to explore

Next /experience

/experience4 entries

Where I've worked

  1. Jun 2026 — Aug 2026

    Skylo

    Cloud

    Software Development Engineer Intern

    • Engineered a centralized MCP Gateway that orchestrates multiple MCP servers behind role-based access control, shared security middleware and structured logging.
    • Designed org-wide AI safety guardrails: system hooks and .mdc rules that stop agents from running destructive commands or unauthorized Git actions.
    MCP gateway architectureCentralized / Policy-enforced / Observable

    MCP gateway architecture. AI and developer clients reach multiple MCP servers only through one centralized gateway. Every request descends the same three checkpoints inside that boundary: authentication, role-based access control, and shared security middleware. Two recorded requests are shown. A read of repository metadata from an IDE clears every checkpoint, crosses the boundary, is routed to an MCP server, and the result is recorded. An unauthorized action from an agent is authenticated but refused by role-based access control: it stops there, never reaching the security middleware or an MCP server, and the block is recorded with the reason RBAC. Structured-log event names are illustrative.

    AI / developer clients

    • IDE
    • Agent
    • Internal
    MCP gatewayOne enforcement boundary
    1. Authenticationpass
    2. RBACblock
    3. Security middleware—
    • Server 01
    • Server 02
    • Server 03

    MCP servers

    Request

    ▲Blocked

    unauthorized action

    Agent→block/RBAC

    Decision trace

    1. [01]request.received
    2. [02]authentication.checked
    3. [03]rbac.denied
    4. [04]block.recorded

    Recorded decisions

    • IDEread repository metadata

      allow/MCP Server 03

    • Agentunauthorized action

      block/RBAC

    • Centralized MCP gatewayMultiple MCP servers
    • Authentication / RBACCentralized access control
    • Structured loggingCentralized observability
    • Event names illustrative

  2. Nov 2025 — Mar 2026

    Samsung R&D

    Researcher & Developer

    • Implemented a 3D Gaussian Splatting pipeline reconstructing scenes from monocular, uncalibrated images, with 43% better fidelity than the COLMAP baseline.
    • Built depth estimation and camera pose pipelines on DepthAnything and DUSt3R, enabling calibration-free reconstruction and cutting manual setup time by 60%.
    Reconstruction pipelineCalibration-free

    Reconstruction pipeline: a single uncalibrated image is turned into a depth field with DepthAnything, camera pose is recovered with DUSt3R, the scene is reconstructed as a sparse point set, and that set is resolved into 3D Gaussian splats.

    1. Input

      Monocular

    2. Depth

      DepthAnything

    3. Camera pose

      DUSt3R

    4. Reconstruction

    5. Gaussian splat

      3DGS

    • +43%Reconstruction fidelityvs. COLMAP baseline
    • -60%Manual setup time

  3. Sep 2025 — Nov 2025

    NeoSapien

    Business & Product Associate

    • Ran LDD and FDD workstreams through a seed round, reviewing 50+ documents and synthesizing findings for investor decks.
    • Analyzed three product verticals and surfaced the growth opportunities that shaped the Q4 roadmap.

  4. Nov 2024 — Feb 2025

    HiDevs

    AI Engineer Intern

    • Architected and deployed a production RAG system on LangChain and FAISS, with 42% lower query latency and 35% better answer relevance.
    • Built document indexing pipelines with chunking and metadata filtering, serving 100+ documents at sub-second retrieval through FastAPI.
    • Co-developed an AI interviewer for automated candidate screening, improving response coherence by 30%.
/projects3 systems

Selected work

Autonomous software engineering system

AgentForge / AutoDev

A multi-stage pipeline that carries software from requirements through architecture, code generation, static checks, QA and real test execution, with human review gates and evidence-driven retry loops between the stages. Typed Pydantic contracts, isolated run workspaces, structured logging, provider fallback and token-budget enforcement keep long runs predictable; failure-aware retry, degrade and abort strategies decide what happens when a stage does not pass. A FastAPI service with SQLite persistence and a server-sent event stream drives the run, executes and lints the generated code, and verifies it against an automated test suite.

  • Python
  • LangGraph
  • FastAPI
  • Pydantic
  • SQLite
  • Pytest

Revenue recovery and auditability system

RecoverOS

A revenue recovery agent, built for the Razorpay Buildathon, that can prove what it did, what it spent and why it stopped. A deterministic policy engine, not the model, owns every spend decision, governed by stopping rules, a consent registry, retry caps, a cost ceiling, expected-value checks and quiet hours. Every outcome is committed to a SHA-256 hash-chained, append-only ledger with tamper detection, and refusals are recorded exactly like actions are. Gemini diagnoses unrecognised errors, reads inbound replies and drafts customer copy, with deterministic fallbacks. Measured against a 20% holdout across 2,000 contacts: +33.9 pp incremental lift, 95% CI +33.0 to +34.9. Razorpay payment links and settlement webhooks run in test mode; customer outcomes and sends are simulated.

  • Razorpay Test Mode
  • Gemini
  • SHA-256 hash chain

Recovery control

▲Tamper detected

Representation of the RecoverOS recovery decision system. A failed payment is ingested and classified, then a deterministic policy engine decides whether to act or refuse. Acting proceeds through a recovery channel; refusing terminates in a recorded reason code. Both outcomes are committed to a SHA-256 hash-chained, append-only audit ledger, and a tamper check demonstration detects a modified entry.

  1. Payment failureok
  2. Ingestok
  3. Classifyok
  4. Policy engineRefuse
  5. Act / channel—
  6. Refuse / reasoncost_ceiling
  7. Audit ledger5 entries

PolicyCost ceiling/Exceeded→Refuse

Ledger / SHA-256 / append-only

▲Tamper detected at 003

  1. 001

    7f3a1c

    payment.failed

  2. 002

    b2e084

    policy.act

  3. 003

    5d9c47

    settle.recorded

  4. 004

    a10f6b

    policy.act

  5. 005

    c48e2d

    policy.refuse

134Automated tests

+33.9 ppIncremental lift95% CI +33.0 to +34.9

Razorpay test mode / outcomes simulated

Reinforcement learning for bug discovery

Software Bug Hunter

A microservice codebase modelled as a connected 18-node, 36-edge Watts–Strogatz graph, where a tabular Q-learning agent learns to walk between modules and test them for bugs. The reward function pays +50 for a critical bug and +10 for a minor one while charging −1 per step, so the policy has to trade coverage against effort; state is discretised to (node, time bucket, tested bucket, already-tested flag), which is what stops the agent oscillating between two high-value modules. The agent is benchmarked against a random-walker baseline on identical episode seeds. Around it sits the MLOps: MLflow experiment tracking, a DVC pipeline driven by a single params file, drift and reward-regression monitoring, versioned policy snapshots with rollback, and a Streamlit dashboard. The published evaluation reports 152.59 average reward against the baseline’s 94.42 (+57.9% reward) at a 97.94% bug-discovery rate over 100 episodes.

  • Python
  • Q-Learning
  • NetworkX
  • MLflow
  • DVC
  • Streamlit
  • Docker

RL environment

✓Episode complete

Reinforcement-learning experiment. A microservice codebase is modelled as a connected 18-node, 36-edge Watts-Strogatz graph of modules such as Auth, Database and API Gateway. A tabular Q-learning agent starts at Auth and moves to a neighbouring module each step, testing it for bugs. Discovering a critical bug pays +50, a minor bug +10, and every step costs -1, so the agent is pushed to find bugs in as few moves as possible. The recorded episode shown runs 22 steps and finds all 6 seeded bugs for a total reward of 118.

  1. Auth
  2. Database
  3. API_Gateway
  4. UI_Frontend
  5. Payment
  6. Logging
  7. Cache
  8. Search
  9. Notifications
  10. UserService
  11. OrderService
  12. Inventory
  13. Analytics
  14. Recommendation
  15. FileStorage
  16. Messaging
  17. AdminPanel
  18. ReportingEngine
Step
22 / 22
Node
Messaging
Event
All bugs found · episode complete
Reward
118
Explored
14 / 18
Bugs found
6 / 6

Reward

+50 critical / +10 minor / -1 step

Policy

Tabular Q-learning / α 0.1 / γ 0.95 / ε 1.0→0.05

RL agent / baseline
Average reward152.59/94.42
Bug discovery rate97.94%/69.82%
+57.9% reward
100 evaluation episodes

Offline experiment / recorded episode

18 nodes / 36 edges / Watts–Strogatz

/about

About

I'm a final-year B.E. student in Artificial Intelligence and Machine Learning at BMS College of Engineering, Bangalore. Most of what I build sits where models meet infrastructure: multi-agent systems, retrieval pipelines, gateways, and the guardrails that make them safe to run in production.

Right now I work on cloud infrastructure at Skylo, where I built a centralized MCP gateway and the org-wide safety rules that constrain what AI agents are allowed to do. Before that I researched calibration-free 3D reconstruction at Samsung R&D, shipped a production RAG system at HiDevs, and worked the product side of a seed round at NeoSapien.

I also lead technical programs for GRADIENT, the AI club at BMSCE, with 100+ members, five hands-on workshops and two hackathons on LLMs, computer vision and agentic workflows.

/contact

Get in touch

Connect with me on LinkedIn. I'm always happy to talk about intelligent systems, developer infrastructure, and reliable backend software.