Hello, I'm
AI Engineer — agentic and LLM pipelines
AI Engineer on the founding team at Colligence Research. I build the product, agentic and LLM pipelines, multi-agent workflows over MCP, a code-generation service that emits deployment-ready applications, backend APIs, and I own the platform it runs on: CI/CD, infrastructure-as-code, observability, incident response and authentication. Most recent proof of both halves at once: a deterministic safety layer under a healthcare LLM, scored by a 62-case labeled evaluation harness that runs as a CI merge gate at recall 1.00. The engineer who ships the agent workflow is the engineer who carries the pager for it.
Requires US visa sponsorship, H-1B, or an F-1/OPT route if studying in the US. Based in Bengaluru, India and open to relocation.
S. Prajwall Narayana
AI Engineer
@Developer1010x
AI Engineer, LLM product engineering and platform ownership
An AI Engineer who ships the product and runs the platform underneath it.
Multi-agent LLM pipelines for ERP automation over MCP, a code-generation service that turns structured input specs into deployment-ready web applications, and the backend APIs behind them. Long-running, multi-step agent workflows that stay on track without a human babysitting them.
The platform those systems run on. I established the DevOps/SRE function at Colligence Research from scratch and lead it: CI/CD, infrastructure-as-code with OpenTofu and Ansible on Linode, Docker, Zitadel for SSO and RBAC, observability, alerting and incident runbooks.
Three peer-reviewed papers, one at Springer ICTIS 2025 in Bangkok and two on IEEE Xplore, covering school administration systems, ResNet50 hyperparameter tuning and a web-server controlled rover with object detection. Two more are under review.
AI and LLM systems and the infrastructure they run on, in equal measure.
Where the work has happened.
Colligence Research • Bengaluru, India
Founding team.
RVCE Center of Excellence • Remote
Ernst & Young India • Bengaluru, India
Academic background.
R V College of Engineering (VTU) • Bengaluru, India
Built end to end, mostly in the open. Metrics are measured, not estimated.
Offline-capable multilingual AI healthcare platform • Jan 2025 – Jun 2025
Symptom checking, X-ray fracture detection, prescription summarization and a wellness chatbot, running fully offline on quantized models, demonstrated on a Raspberry Pi. A deterministic rule-based drug-interaction and crisis-detection safety layer sits underneath and the LLM cannot override it. A 62-case labeled evaluation harness scores those safety layers and runs as a CI merge gate at recall 1.00; building it found a real missed self-harm disclosure. A LangGraph pipeline turns an uploaded medical report into a plain-language explanation in the patient's own language, with a self-verification loop, and requests are routed across local Ollama models. Third place, college final-year showcase.
Private reader and audiobook player for Linux
Reads your own books aloud in a neural voice, with follow-along highlighting and OCR for scanned pages. Everything runs locally; nothing is uploaded. Released as an AppImage, .deb and tarball.
AI-powered CI/CD pipeline generator and monitor
Generates and monitors CI/CD pipelines for GitHub and GitLab repositories.
AI access from a phone
Routes AI access across WhatsApp, Telegram and email so it can be driven from a phone.
Synthetic speech detection
Flags AI-synthesized or cloned speech.
Terminal-based LLM assistant
A lightweight terminal-based assistant for working with LLMs from the command line.
Student resource platform
A student resource platform, adopted across 12+ departments.
Open-source screen casting in Rust
Casts a screen to AirPlay and Miracast receivers with no proprietary software. AirPlay HAP pairing, FairPlay, RTSP negotiation and Wi-Fi Direct implemented from scratch in Rust.
File conversion with multi-hop routing
Finds a conversion path across converters when no direct one exists, using graph search over the registry. Audits which conversions actually work on the current machine instead of advertising formats it cannot handle.
Desktop MP3 tag, artwork and EQ editor
Edits MP3 tags, fetches album art and lyrics, and applies a five-band equaliser across a batch. Non-destructive: EQ is applied only on export. Packaged for the Snap Store.
Merged upstream PR, not my project
A contribution, not a project of mine. Added an expandable price chart to the markets panel of a 330k-line TypeScript intelligence dashboard maintained by Elie Habib, and fixed a dev proxy path bug. Merged as PR #5459.
Compiler plugin measuring control-flow complexity per function • Aug 2026
A Clang plugin that reports the cyclomatic complexity of every function in the file being compiled. It loads into the compiler itself, so it sees the code after preprocessing, with the real include paths and feature flags, rather than reparsing text. It walks the AST instead of building a control-flow graph, which works because M = E − N + 2P collapses to decision points plus one for a single function. It analyzes C and C++ written for GCC as well as Clang, including the GNU extensions real codebases use: statement expressions, the binary conditional with an omitted operand, computed goto, case ranges and __attribute__. Originally a 2024 college project written to understand how a compiler represents code once it stops being text; rebuilt in 2026 for LLVM 18, which meant fixing four counting errors the original had, among them treating a switch as one decision rather than one per case, and folding a lambda's branches into its enclosing function.
Merged upstream bug fix, not my project • Aug 2026
An open-source contribution rather than a project of mine. community.general is Ansible's largest community collection, depended on by a very large number of production automation setups. The snap module crashed with an IndexError instead of reporting a missing snap, because `snap info` reports that condition two different ways: a warning on stdout with return code 0 when several snaps are queried, and an error on stderr with return code 1 when only one is. The parser matched only the first form, so the single-snap case fell through to code that indexed an empty result. Pull request 12570, merged August 2026, handles both forms and fixes the reported issue.
Available for select freelance projects, I build DevOps pipelines and RAG systems for real workloads, with the evaluation to prove they work.
Retrieval-augmented systems over your own corpus, hybrid retrieval, reranking, and a labeled evaluation harness wired into CI so quality is measured, not hoped for. The same discipline behind Arogya-Sathi's 62-case, CI-gated safety eval (recall 1.00).
The platform your product runs on, CI/CD, infrastructure-as-code, containerization, observability and incident runbooks. Built the DevOps/SRE function from scratch at a founding-team startup and carry the pager for it.
Multi-agent and LLM systems over MCP with LangGraph/LangChain, kept on-device with local models where the data is sensitive, and shipped as a monitored service, not a demo.
Peer-reviewed work.
Springer ICTIS 2025, Bangkok, ICT for Intelligent Systems, Vol. 9
IEEE Xplore, November 2024
Read the paperIEEE Xplore, November 2024
Read the paperTwo further papers are under review.
The fastest way to reach me is email.
Bengaluru, India • Open to relocation
Requires US visa sponsorship, H-1B, or an F-1/OPT route if studying in the US. Based in Bengaluru, India and open to relocation.