Vellore Institute of Technology
First Class with Distinction · 8.89 / 10.
Animesh Seth
Software engineer working in GenAI
sethanimesh@hotmail.com · GitHub · LinkedIn
Profile
I’m Animesh, a software engineer in the GenAI team at Tally Solutions. My work spans AI applications, retrieval, agent orchestration, evaluation, and inference infrastructure.
My personal projects focus on assistive speech, vocabulary learning, and robotics. I work across model adaptation, backend systems, and interfaces, with an emphasis on context, reliability, and user control.
Education
Experience
Tally Solutions
January 2025–presentDeveloping enterprise GenAI applications for product support and software engineering. Contributions span context-aware retrieval, agent workflows, model evaluation, and inference optimisation, alongside shared infrastructure for memory, safety, recovery, and observability.
Proteus Technologies
November–December 2023Worked on computer vision for inventory workflows, combining edge-model optimisation with CLIP-based classification. Quantised and benchmarked vision models and developed a pipeline for identifying industrial chemical labels.
Selected work
TARA — Multimodal Support Assistant
I contributed multimodal ingestion, region-aware retrieval, and human-agent handoff logic to TARA, an existing support product. My work focused on interpreting screenshots, audio, and video, retrieving relevant guidance, and preserving conversation context during escalation.
Automated Code Review
I built an AI-assisted code-review platform integrating repository analysis, static analysis, security checks, and pull-request workflows. The review architecture supports source investigation and evidence-backed findings.
AI Evaluation Framework
I built an internal evaluation framework for AI responses, using completeness and relevance criteria with pass, fail, and escalation outcomes. Ambiguous cases are routed to engineers for review.
Unified AI Runtime
I built a shared backend runtime for product-support and developer copilots, with common capabilities for memory, safety, failure recovery, and tracing. The applications remain in development and QA.
Enterprise Agentic SDLC Framework
I designed and implemented a human-in-the-loop workflow for requirements qualification. It gathers stakeholder clarifications, maintains a working requirements document, and pauses for input before proceeding.
Echora
I developed adapted speech recognition and an assistive communication interface with editable wording and visible alternative interpretations. The project explores how recognition and generated wording can support communication while retaining speaker control.
Code and researchRecallX
I built a source-aware vocabulary application with flashcards, shared review history, FSRS scheduling, and recoverable offline reviews. Self-rated practice is usable; automated semantic assessment remains experimental.
Code and researchAutonomous Assistive Robot
I designed and tested a Jetson and LEGO EV3 robot with a single tilting camera for selected-person search and bounded movement. The work includes identity continuity, camera coordination, and handling delayed perception results.
Code and researchResearch and writing
Audio-Visual Emotion Recognition
I developed a cross-modal Transformer combining facial and speech representations to study emotion recognition in children with autism spectrum disorder and intellectual disabilities.
The Hidden Cost of Structured Output: How It Can Suppress LLM Reasoning
A technical analysis of how structured-output constraints can interrupt repository investigation, and an architecture that separates tool-assisted analysis from output formatting.
Toolkit
Python, C++, SQL, PyTorch, FastAPI, LangGraph, Google ADK, MCP, vLLM, Docker, ROS 2, OpenCV, TensorRT.