ScopeLock: An Offline, NDA-Safe Proposal Scoper for Freelancers Using Local Gemma Models

·1 views

A developer built ScopeLock, an open-source tool that helps freelancers scope client proposals without sending confidential briefs to cloud AI services. The tool runs open-weights Google Gemma models locally through Ollama and ingests PDF, Markdown, or plaintext briefs entirely on the user's machine. It extracts requirements anchored to verbatim quotes, computes pricing and milestones with deterministic code rather than the language model, and flags ambiguous client language with protective contract clauses. An automated canary test is included to verify that no client content leaves the machine, and the project is released under the Apache-2.0 License. The hosted web demo replays traces from local inference rather than running cloud inference.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

InvoiceMate AI: A Local Open-Model Tool for Organizing Freelancer Expenses

InvoiceMate AI is a small expense organizer aimed at freelancers and tiny businesses that need to turn piles of invoices and receipts into clean records. Users paste or upload invoice text, and a local open-weight model running through Ollama extracts the vendor, date, amount, category, and description for review. Extracted entries are shown in an editable form before saving, and saved expenses appear on a dashboard with totals, a category chart, and a CSV export. The developer emphasizes privacy, since invoice text stays on the user's machine rather than going to a hosted AI service, though the project is an early MVP that holds data only in the browser session and does not support scanned-image OCR.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Keeping Original MRI Slices Distinguishable From AI-Generated Interpolations

A developer's exploratory MRI visualization project combined Python, AI-generated intermediate images, background transparency, and Blender. The article argues that generated slices can become indistinguishable from source images, making it harder to trace what came from the original data. It proposes keeping imported images in an immutable collection and writing derived results, masks, and generated layers separately with references to their sources. A manifest recording identifiers, ordering, spatial metadata, and processing settings is suggested, along with clearly marking generated images at creation. The piece also cautions that evenly stacked images do not establish physically accurate spacing and that generative interpolation is not the same as acquiring new observations.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Tenant-Filtered Vector Search Needs Matching Embedding Models and Chunk-Level Freshness

The article argues that retrieval systems should embed user queries with the same model used for indexed help-center content and apply tenant constraints within the vector query itself. Filtering after a broad top-k search is described as flawed, since other tenants' chunks enter the candidate pool and displace permitted results. It recommends chunking content into answer-sized units and tracking chunk ID, tenant ID, source revision, and embedding model ID so changed material can be re-indexed selectively. The author also advises keeping telemetry lean, avoiding high-cardinality labels such as raw query text, full vectors, and tenant or chunk IDs. The piece frames these as correctness decisions that apply regardless of which vector database a team chooses.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

RTX 3090 vs L40S: Relative Speed Depends on Batch Size, Study Finds

A developer measured an RTX 3090 against an L40S on rented machines using the same script, testing 7B fp16 models. The 3090 matched the L40S at batch size 1 (1.09x) but reached only 0.46x its speed at batch size 128. The author concluded there is no single speed figure, since relative performance depends on how many requests run concurrently. Cost was also factored in, with the 3090 rented at $0.414/h and the L40S at $0.826/h. The author noted that a different 3090 unit tested earlier was slower across all settings, so hardware should be measured per unit. All 192 runs and the script are published on GitHub.

0 comments Read more at DEV Community