Agentic AI Engineer

https://drone-port.in/; Chennai District; On-site; INR 1,200,000 - 1,800,000; Full time

Job summary

drone-port.in seeks a software engineer to build Tamil-language assist services for farmers with limited literacy. You will own the end-to-end assist service, including Tamil chat, a read-only copilot, and the evaluation and guardrail layers they run behind.

You’ll design and ship an evaluation set of Tamil questions and enforce strict usage caps to prevent runaway costs. The role requires shipping an LLM feature to real users, with strong emphasis on cost discipline and listening to farmers to

Qualifications

• 3–5 years in software with a shipped LLM feature to real users
• Experience designing prompts and integrating tools for AI agents
• Experience building evaluations, guardrails, and logging
• Ability to explain model trade-offs, batching and caching without vendor bias
• Interest in Tamil language and willingness to work with farmers
• Able to listen to farmers and understand non-straightforward user questions

Responsibilities

• Own the assist service: Tamil chat help and read-only operations copilot
• Develop an evaluation set of 100 Tamil questions with published pass rate
• Enforce guardrails, per-generation logging, and caps at the gateway
• Set a budget alarm at 60% of monthly allocation by day fifteen
• Measure voice booking assist with real farmers in Erode prior to release
• Ensure generated text does not reveal numbers not computed by the system

Skills

LLM feature shipped

Prompt design

Tool-use design

Agent loops

Evaluations & guardrails

Cost discipline

Tamil interest

Tools

Budget tools

Job description

The role

You build Tamil-language assistance that a farmer with limited literacy can actually use, with evaluations and guardrails from day one rather than added after the first embarrassment. You own the whole assist service, and you have the architecture’s backing to keep AI out of money and safety paths.

Key responsibilities

• Own the assist service: Tamil chat help, a read-only operations copilot, document generation, and the evaluation and guardrail layer they run behind
• Build an evaluation set of at least a hundred Tamil questions with a published pass rate, run in the pipeline and not by hand
• Enforce guardrails, per-generation logging and hard per-user and per-day caps at the gateway rather than in application code
• Set a budget alarm at sixty per cent of the monthly allocation by day fifteen - the classic failure here is not a wrong answer, it is a bill
• Measure voice booking assist on real farmers in Erode before it enters the release, and never let it auto-submit anything
• Keep AI permanently out of every money, dispatch and safety path - generated text never contains a number the system did not compute

Eligibility

• Three to five years in software, with at least one LLM feature shipped to real users rather than demonstrated
• Prompt and tool-use design, and agent loops
• Evaluations and guardrails you actually wrote and ran, with a story about a bad answer you caught before a user did
• Cost discipline against a real budget - you can explain model trade-offs, batching and caching without naming a favourite vendor by reflex
• An interest in Tamil; reading it is enough to start
• Willing to sit with farmers and listen to what they actually ask, which is never what was expected

What you get

• A genuinely hard language problem with real users at the end of it
• Told plainly before you join where this work sits if the programme needs time - you will not discover that in November

How the process runs

• We read it ourselves. Every application is read.
• If we want to meet you, we call you.