I build production multi-agent systems and fine-tuned SLMs, turning intractable enterprise impasses into measurable business value.

AI Systems Engineer
+
Builder & Systems Thinker
+
Hyderabad, IND

Biyyala Saathvik

Systems engineering

I take an ambiguous problem from whiteboard logic to shipped code and stay on it until it runs on its own.

Foundations & tools

I build the thing that makes everyone else faster, then keep it alive and reliable long after launch.

Research that ships

I run the experiments myself and tie every evaluation to a real system that ships.

Selected Systems

Empathetic Financial Guidance Agent

Architected and deployed a production LangGraph multi-agent system with Human-in-the-Loop (HITL) orchestration in 3 months for a Tier-1 UK bank following a 9-month client impasse, earning the Best Executive AI-Led Service Delivery Award.

Project Timeline: 3 Months

Recognition: TCS Blitz 2026 Winner

Role: AI Systems Engineer

Core Stack

Proteus — Agentic AI Workspace

Engineered an open-source agentic workspace that closes the gap between ChatGPT's polish and a homelab's freedom. Solved tool-calling token bloat with semantic top-k tool indexing across large MCP registries.

Features ~60 FastAPI route modules, a responsive PWA, 4 execution sandboxes, persistent memory via ChromaDB + ONNX, and a git-style CLI with ~20 subcommands.

Architecture: Solo System Architecture

Status: Open Source (GitHub)

Role: Creator & Lead Engineer

Core Stack

Financial Crime Investigation Pipeline

Transformed labor-intensive anti-money laundering investigations into an auditable multi-agent workflow where specialized agents investigate typologies, cross-reference evidence, and draft regulatory filings.

Engineered an autonomous 9-agent AML orchestration pipeline automating end-to-end case lifecycles, cross-source typology analysis, and rubric-based iterative refinement.

Domain: BFSI AML Compliance

Scale: 9-Agent Architecture

Role: AI Systems Engineer

Core Stack

Systems & Research

From zero-shot computer vision to domain-adapted language models, exploring how AI systems scale with precision and autonomy

Legacy Code Modernization SLM

Domain-adapted 7B SLM trained on a curated 7B-token corpus with synthetic data generation, CPT, and SFT for bidirectional code engineering across 8 languages.

ACE Framework & Clicks2Converse

Engineered the Adaptive Conversational Experience core architecture (collapsing 20+ navigation clicks into a single conversational turn; 95% step reduction), cited in Celent's Global Flash Report.

Real-Time Traffic Sign Distance Estimation

Real-time computer vision pipeline pairing YOLOv8 object detection with Depth-pro zero-shot monocular depth estimation for edge autonomous vehicle hardware.

MIRA & Unified Asset Aggregator

Agentic mortgage advisory engine pairing domain-aware RAG with an asset consolidation pipeline tracking real-time wealth holdings across heterogeneous systems.

About Me

I'm Saathvik. I engineer agentic architectures and domain-adapted models, mostly focused on high-stakes systems where hallucination isn't an option and reliability is everything.

I've always been drawn to systems that solve their own problems. When I'm not running fine-tuning jobs or refining agent prompts, I love exploring emerging AI research, working on side projects like Proteus, and enjoying quiet evenings with good music and great conversations.