Website:
craftifai.com
Job details:
Company: CraftifAI
Work arrangement: Full-time, work from office
Role focus: 70% hands-on research and engineering; 30% technical leadership and ownership
Preferred joining timeline: Candidates who can join in fewer than 45 days are preferred.
About CraftifAI
CraftifAI is building Orbit, a platform that automates engineering workflows across embedded software, Edge AI, and FPGA development. Our work connects requirements, domain knowledge, code generation, testing, validation, and deployment to help engineering teams build working products.
We are looking for a research-driven engineer who can strengthen the LLMs and agentic systems behind these workflows, with a particular focus on efficient on-premises deployment and reliable code generation.
About the Role
You will lead the research and implementation of domain-specific LLM solutions and agentic systems for engineering applications. This includes fine-tuning models, reducing their deployment footprint while preserving measured task quality, designing knowledge architectures, building agent harnesses, and establishing rigorous product benchmarks.
This is a hands-on technical leadership role. Approximately 70% of your time will involve coding, experiments, model optimization, evaluation, and product integration. The remaining 30% will involve architecture decisions, research direction, mentoring, and ownership of technical outcomes.
You should be comfortable moving from a research hypothesis to an implementation, measuring its impact, understanding its failures, and improving it until it works reliably within the product.
What You Will Own
1. LLM Fine-Tuning and Efficient On-Premises Deployment
- Adapt LLMs for engineering tasks, including code generation, tool use, debugging, and reasoning over domain-specific information.
- Build data preparation and fine-tuning workflows using suitable approaches such as supervised fine-tuning, LoRA, QLoRA, and preference optimization.
- Apply and evaluate compression techniques such as quantization, knowledge distillation, and pruning to reduce model size and memory requirements.
- Measure trade-offs between task quality, model footprint, inference latency, throughput, and infrastructure cost on target deployment hardware.
2. Agentic Systems for Code Generation
- Design and implement agent harnesses for task planning, repository exploration, context management, code generation, tool execution, debugging, and iterative correction.
- Build reliable execution flows with clear state management, failure recovery, stopping conditions, and observable results.
- Improve the ability of agents to complete engineering tasks across multiple files, dependencies, tools, and execution environments.
- Own the integration of these capabilities into CraftifAI’s product workflows.
3. Domain Knowledge and Context Architecture
- Design knowledge systems that connect source code, requirements, architecture documents, hardware documentation, SDKs, APIs, and previous engineering outcomes.
- Develop effective retrieval and context construction using approaches such as hybrid search, reranking, structured knowledge, and repository-aware retrieval.
- Decide when information belongs in model training, retrieval, structured tools, or persistent agent memory.
- Keep retrieved information traceable to its source, version, and applicable hardware or software configuration.
4. Code Quality and Automated Verification
- Build and integrate static analysis tools, compiler diagnostics, linters, code structure checks, and security checks into agent workflows.
- Combine these checks with automated builds, unit tests, and controlled execution to evaluate generated code.
- Convert tool findings into actionable feedback that agents can use to diagnose and correct their outputs.
- Establish quality gates and regression checks that provide objective evidence of improvement.
5. Low-Latency System Design
- Profile and optimize the complete workflow, including retrieval, context preparation, model inference, tool calls, and correction cycles.
- Improve serving, caching, batching, context efficiency, and execution routing where they produce measurable benefits.
- Design for practical on-premises constraints, including available compute, GPU memory, concurrency, and restricted external connectivity.
- Optimize time to a verified result alongside model response time.
6. Reasoning and Learning from Field Use
- Improve agent planning, tool selection, debugging, and recovery through systematic failure analysis and controlled experiments.
- Turn field failures and user feedback into curated datasets, evaluation cases, improved knowledge, and better agent behavior.
- Evaluate suitable methods for improving reasoning and tool use, including fine-tuning, preference learning, and reinforcement learning where appropriate.
- Validate improvements against regression benchmarks before release, within customer data and deployment constraints.
7. Product Benchmarking and Technical Leadership
- Own reproducible benchmarks at both model and complete product workflow levels.
- Measure task completion, functional correctness, static analysis findings, test results, human intervention, latency, memory use, and compute cost.
- Use held-out tasks, baseline comparisons, ablation studies, and error analysis to explain which changes improve the product and why.
- Define research priorities, lead architecture and code reviews, mentor engineers, and take ownership from experimentation through product integration.
Required Qualifications and Experience
- Minimum qualification: MS by Research from a Tier-1 institution in India or abroad, in Computer Science, AI/ML, or a closely related field. A PhD is preferred.
- Strong practical experience fine-tuning LLMs and deploying them on private or on-premises infrastructure.
- Demonstrated experience reducing model size or inference memory requirements, supported by measured quality and performance results.
- Deep hands-on experience building agentic systems for code generation, including tool integration, execution feedback, and iterative repair.
- Strong understanding of domain knowledge architecture, retrieval systems, context management, and their interaction with model behavior.
- Strong Python and software engineering skills, with practical experience using frameworks such as PyTorch, Transformers, PEFT, or equivalent tools.
- Experience integrating static analysis, build systems, automated testing, or other programmatic checks into code-generation workflows.
- Strong research discipline: clear hypotheses, controlled experiments, reproducible results, and careful interpretation of failures.
- Ability to make architecture decisions, communicate technical trade-offs, and personally implement critical parts of the solution.
Additional Experience We Value
- Research publications or open-source contributions in efficient LLMs, code models, reasoning, agentic systems, or program analysis.
- Experience with C/C++, embedded software, compilers, firmware, or hardware-aware code generation.
- Experience with model serving, inference profiling, distributed training, or resource-constrained deployments.
- Experience turning research prototypes into maintained production capabilities.
What to Include in Your Application
Please share your CV and relevant publications, repositories, or project examples. Where possible, include:
- An LLM fine-tuning or compression project with measured quality, memory, and latency results.
- An agentic code-generation system you personally built, including its tool integrations and evaluation approach.
- A brief explanation of your individual contribution and the most significant technical problem you solved.
- Your current notice period and availability to work from the office.
Click on Apply to know more.
