GigLift
A prototype of a small freelance task platform for students in India. It uses large models to convert spoken Hinglish requests into structured task briefs, with a division of labor in which rules extract budgets and deadlines while the model interprets the language.
Timeline (last 90 days)
GigLift is a small task platform for students that turns messy, colloquial requests (such as Hinglish messages on WhatsApp) into formal task briefs containing deliverables, skill requirements, budget, and deadline.
GigLift's tech stack is Django with HTMX and SQLite, deployed on Render's free tier.
All of GigLift's AI logic is concentrated in a single file, core/llm.py, using only the standard library and introducing no SDK.
GigLift's design decision is that budget and deadline are not left to the model to judge but are extracted directly from the text by the rule set in core/extract.py, and the model has no authority to modify them.
GigLift's rule set can handle English, Hinglish, and Devanagari numerals, such as 1500 tak, ₹१२००, parso (the day after tomorrow), agle hafte (next week), and Friday tak.
If the deliverables generated by the model contain numbers that do not exist in the client's message, that line is simply discarded.
In a held-out evaluation on 32 briefs, the new rules read 22 of the explicit budgets (25) correctly, versus 11 for the old regex; and 25 of the explicit deadlines (27) correctly, versus 8 for the old regex.
In a held-out evaluation on 32 briefs, the new rules fabricated nothing at all for the 7 briefs with no budget given (the old regex fabricated all 7) and fabricated nothing at all for the 5 briefs with no deadline given (the old regex fabricated all 5).
Running the full pipeline with real Gemma on 32 briefs: 22 of 25 budgets read correctly, 25 of 27 deadlines read correctly, zero errors, zero fabrication.
GigLift's evaluation has limitations: the rules and the test briefs were both created by the same person, so the held-out results are somewhat optimistic; the evaluation only measures numbers and has not yet assessed the quality of the open questions, skills, and deliverables generated by Gemma.
4 more timeline events
All Facts (14)
GigLift is a small task platform for students that turns messy, colloquial requests (such as Hinglish messages on WhatsApp) into formal task briefs containing deliverables, skill requirements, budget, and deadline.
50%UnverifiedGigLift's tech stack is Django with HTMX and SQLite, deployed on Render's free tier.
50%UnverifiedAll of GigLift's AI logic is concentrated in a single file, core/llm.py, using only the standard library and introducing no SDK.
50%UnverifiedGigLift's design decision is that budget and deadline are not left to the model to judge but are extracted directly from the text by the rule set in core/extract.py, and the model has no authority to modify them.
50%UnverifiedGigLift's rule set can handle English, Hinglish, and Devanagari numerals, such as 1500 tak, ₹१२००, parso (the day after tomorrow), agle hafte (next week), and Friday tak.
50%UnverifiedIf the deliverables generated by the model contain numbers that do not exist in the client's message, that line is simply discarded.
50%UnverifiedIn a held-out evaluation on 32 briefs, the new rules read 22 of the explicit budgets (25) correctly, versus 11 for the old regex; and 25 of the explicit deadlines (27) correctly, versus 8 for the old regex.
50%UnverifiedIn a held-out evaluation on 32 briefs, the new rules fabricated nothing at all for the 7 briefs with no budget given (the old regex fabricated all 7) and fabricated nothing at all for the 5 briefs with no deadline given (the old regex fabricated all 5).
50%UnverifiedRunning the full pipeline with real Gemma on 32 briefs: 22 of 25 budgets read correctly, 25 of 27 deadlines read correctly, zero errors, zero fabrication.
50%UnverifiedGigLift's evaluation has limitations: the rules and the test briefs were both created by the same person, so the held-out results are somewhat optimistic; the evaluation only measures numbers and has not yet assessed the quality of the open questions, skills, and deliverables generated by Gemma.
50%UnverifiedGigLift itself does not make money; no real customer has used it yet, and no one has paid through it; it is a deployed, runnable prototype rather than a mature product.
50%UnverifiedIn GigLift, students and shop owners exchange phone numbers only after both sides select a match, and payment is completed offline.
50%UnverifiedGigLift uses the MIT license, its code is public, and its evaluation is reproducible.
50%UnverifiedOn security, GigLift implements CSP with nonce, per-session and per-IP rate limiting, ownership checks for edit and hire operations, and has 24 Django tests plus 7 rule tests.
50%