The posting
About Impact Analytics
Impact Analytics is an agentic-first AI software company transforming retail merchandising through cutting-edge AI, LLMs, and Generative AI technologies. As a fast-growing Series D company with deployments across five continents, it is building both industry-leading merchandising solutions and foundational AI agents that are redefining how retail decisions are made. What makes Impact Analytics unique is its combination of deep retail domain expertise, strong innovation culture, and global presence. It is one of the few India-born AI companies recognized globally by organizations like Fortune, Gartner, and the Inc. 5000.For candidates looking to work on next-generation AI products with global scale and real-world impact, Impact Analytics is an exciting place to build your career. Here’s a link to our website: www.impactanalytics.co.
Overview
The AI Senior Software Engineer builds and ships production-grade AI features within a defined domain of the company’s applied AI platform spanning our GenAI-powered reasoning and analytics platform, our agentic AI authoring and runtime platform, and emerging multimodal / computer vision workloads. You take a prioritised backlog and the architectural guidelines set by the AI Lead and AI Architect and turn them into shipped, well-tested AI systems. This is a hands-on individual-contributor role. You write and review production code, contribute to design decisions within your squad, and may informally guide one or two junior engineers. You do not own a domain’s architecture or lead a squad you implement and evolve features within it, flag issues upward, and own the quality of your own work.
Key Responsibilities
1. Platform Implementation
• Implement and evolve features across the GenAI reasoning and agentic AI platform in alignment with the reference architecture defined by the AI Architect.
• Contribute to the technical design of your squad’s workstreams — data flow, API contracts, component boundaries, and integration points with upstream/downstream systems.
• Write and review production-quality Python code; follow code standards, participate in PR reviews, and help maintain a clean, well-tested codebase.
• Contribute input to build-vs-buy-vs-open-source evaluations for your features; provide trade-off analysis to the AI Lead / AI Architect, who own the final decision.
• Contribute to Architecture Decision Records (ADRs) for decisions within your scope; review ADRs as requested.
2. Agentic AI Systems
• Build and ship agent workflows (ReAct, planner–executor, reflection, human-in-the-loop checkpoints) using the chosen orchestration framework (LangGraph / LlamaIndex Workflows / Semantic Kernel or equivalent).
• Implement tool and MCP (Model Context Protocol) integrations for connecting agents to enterprise systems (CRM, ITSM, REST/GraphQL APIs) with scoped auth (OAuth2/OBO), sandboxing, and rate limiting.
• Contribute to agent evaluation coverage: task-completion benchmarks, tool-call success rates, trajectory evals, and cost-per-task regression tests.
• Implement agent memory patterns (short-term conversation context, long-term episodic/semantic memory) following TTL and retrieval policies defined by the AI Lead / AI Architect.
3. GenAI & RAG Systems
• Implement and iterate on RAG pipelines: hybrid retrieval (BM25 + dense + rerankers), query decomposition, GraphRAG for knowledge-graph workloads, and citation grounding.
• Execute on the fine-tune vs. prompt vs. in-context-learning decision set by the AI Lead / AI Architect; carry out fine-tuning runs (LoRA/QLoRA/PEFT) with guidance when required.
• Apply context engineering patterns: prompt contracts, structured outputs (JSON Schema, Pydantic, function-calling), constrained decoding, output validators, and fallback chains.
• Implement grounding and hallucination-mitigation mechanisms: retrieval confidence scoring, “I don’t know” paths, claim-level validators.
4. Multimodal & Computer Vision (Contextual Scope)
• Where the product roadmap requires it, build VLM-first pipelines (GPT-4V class, Gemini, Qwen-VL) and integrate CV outputs (OCR, document understanding, layout parsing) as tools inside the agentic platform.
• Run benchmarks on open-weight VLMs (Qwen-VL, InternVL) vs. closed VLM APIs for your squad’s use cases; report findings with cost-latency-accuracy data.
5. LLMOps, Evaluation & Observability
• Instrument telemetry for the features you build: token cost, latency p50/p95, tool-call success rates, retrieval hit rates, cache hit rates — per-tenant where applicable.
• Maintain continuous evaluation coverage: hallucination, faithfulness, jailbreak resistance, and drift detection for the features you build; treat prompts as code (versioned, reviewed, rolled back).
• Use and extend the LLMOps toolchain (LangSmith / Langfuse / Arize / Ragas / DeepEval /Promptfoo) as established by the platform team.
6. Responsible AI, Security & Governance
• Implement AI security controls for the features you build: prompt-injection defense, PII detection and redaction, RBAC on agent tools, and secrets handling for tool auth.
• Ensure the features you build meet compliance requirements (SOC 2 Type II, GDPR, DPDP) as defined by the platform governance team.
• Apply responsible AI guardrails: topic filters, PII egress controls, usage-policy enforcement at gateway level.
7. Inference Economics & Performance
• Profile and optimise inference for your workloads: batching, KV-cache utilisation, speculative decoding, and quantisation (AWQ, GPTQ, INT4) under guidance from the AI Lead / AI Architect.
• Provide SLM-vs-LLM routing input for individual capabilities (e.g., Phi-3.5 / Qwen-2.5-7B vs. frontier API) with supporting cost-latency-quality data.
• Implement and instrument caching layers (prompt cache, semantic cache, retrieval cache) with invalidation policies.
8. Collaboration & Knowledge Sharing
• Share knowledge through design reviews, PR feedback, and pair programming; support junior engineers as needed.
• Participate in squad design sessions; surface cross-cutting concerns to the AI Lead, AI Architect, and Engineering Manager proactively.
• Partner with Product, SRE, and Security on sprint-level execution, incident triage, and feature flag / release decisions.
• Participate as a technical interviewer for engineering roles on the AI team.
Required Qualifications
• 5+ years of software engineering experience, with 2+ years shipping production AI/ML systems and hands-on experience delivering LLM-powered applications (RAG, agents, fine-tuning).
• Hands-on proficiency in at least one agent orchestration framework (LangGraph, Semantic Kernel, CrewAI, AutoGen, DSPy).
• Experience with vector stores (pgvector, Pinecone, Weaviate, Milvus, Qdrant), embedding models (open + proprietary), and hybrid retrieval with rerankers.
• Working knowledge of fine-tuning (LoRA/QLoRA/PEFT) and inference optimisation (vLLM, quantisation, batching) — implementation-level, not just conceptual.
• Solid MLOps / LLMOps foundations: experiment tracking, CI/CD for ML, prompt versioning, and evaluation pipelines (LLM-as-judge, RAG triad metrics, adversarial suites).
• Cloud experience on at least one of AWS (Bedrock, SageMaker), Azure (AI Foundry, Azure OpenAI), or GCP (Vertex AI) — including cost management, not just API usage.
• Python proficiency at senior-engineer level; working knowledge of PyTorch and the HuggingFace stack.
• Comfortable with LLM evaluation methodologies: faithfulness / answer-relevance / context- relevance, agent trajectory evals, and human-in-the-loop eval workflows.
• Strong communication — can write a clear design doc that both a product manager and an engineer find useful, and can explain technical trade-offs in sprint planning.
Preferred Qualifications
• Experience building features for a multi-tenant GenAI SaaS product (not a pilot or internal tool).
• Hands-on with MCP (Model Context Protocol), A2A, or equivalent agent-tool interoperability standards.
• Exposure to VLMs / multimodal systems (GPT-4V, Gemini, Qwen-VL, InternVL, LLaVA) and VLM fine-tuning.
• Experience with small language models (SLMs) and on-prem / edge deployment (Phi, Gemma, Qwen, Llama 3.1-8B class).
• Familiarity with graph databases (Neo4j, Neptune) for GraphRAG and knowledge-graph workloads.
• Public contributions to AI open-source, technical blog posts, or conference talks are a plus.
• Graduate degree (MS) in CS, ML, or related — shipped systems outweigh credentials.
What we offer
• An opportunity to be part of some of the best enterprise SaaS products to be built out of India.
• Opportunities to quench your thirst for problem-solving, experimenting, learning, and implementing innovative solutions.
• A flat, collegial work environment, with a work hard, play hard attitude.
• A platform for rapid growth if you are willing to try new things without fear of failure.
• Remuneration with best-in-class industry standards with generous health insurance cover
Some of our accolades include:
• Ranked as one of America's Fastest-Growing Companies by Financial Times for five consecutive years: 2020-2024.
• Ranked as one of America's Fastest-Growing Private Companies by Inc. 5000 for seven consecutive years: 2018-2024.
• Voted #1 by more than 300 retailers worldwide in the RIS Software LeaderBoard 2024 report.
• Ranked #72 in America’s Most Innovative Companies list in 2023—by Fortune—alongside companies like Microsoft, Tesla, Apple, IBM, etc.
• Forged a strategic partnership with Google to equip retailers with cutting-edge generative AI tools.
• Recognized in multiple Gartner reports, including Market Guides and Hype Cycle, spanning assortments, merchandising, forecasting, algorithmic retailing, and Unified Price, Promotion, and Markdown Optimization Applications.
